but it's much closer to what actual self-identified utilitarians do, most of the time, than any version of a literal utilitarian calculus.
mostly, for instance, utilitarian EAs do things that are pretty much in character and congenial, given their personalities, socioeconomic classes, professions, and so on. within the range of actions that are in their comfort zone, they may sometimes try to do the utilitarian-best thing, but they're normal humans, not abstract optimizers.
this why for instance I am unconvinced by claims like "EAs are dangerous because they're utilitarians who'll do Anything, even terrorism/crimes, for the Greater Good." No; they're mostly young urban professionals who do things typical of young urban professionals.
radicalization into the belief that it's necessary to do terrible things For The Greater Good is a real thing, but you have to be in a social environment that specifically normalizes those terrible things -- or to be socially alienated, like a Ted Kaczynski type.
and it's probably more correct.
I want someone to think in terms of the "greatest good for the greatest number" when they are an administrator making decisions in a hospital, for instance. I want people to try to benefit more people, to a greater degree, in contexts where the kind of people affected, the kinds of risks and benefits being evaluated, and the kinds of actions within scope, are already determined and deemed acceptable (generally by something other than utilitarianism.)
I don't want people to consider "what produces the greatest good for the greatest number" regarding every action they take, and of course nobody does this in real life.
https://distill.pub/2020/growing-ca/ neural cellular automata: "Typical CA models represent cell states with a set of discrete values, although variants using vectors of continuous values exist. The use of continuous values has the virtue of allowing the update rule to be a differentiable function of the cell’s neighbourhood’s states".
these look lifelike and also a lot like the max-activating of CNN neurons, which makes sense. (maximizing the activation of a CNN neuron via gradient descent is the result of repeatedly updating a differentiable function of each cell's neighborhood's states)